A decades-old PHP flaw in 70% of the web, 17 protein binders, and AI reviews that beat the experts

In the three days to October 23, 2026, eight items from the AI and tech accounts this channel follows were worth carrying. AI found two critical bugs in the PHP engine that runs much of the web, and the patch landed. Claude designed protein binders against scorpion venom. A Harvard paper reported AI reviews beating expert human ones. The window runs three days because the editions of October 21 and October 22 were not published.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked. One of those accounts, Benedict Evans, can no longer be read on X, so its posts are absent from this edition.

AI found two critical bugs in the PHP engine that runs 70% of the web Wiz Research's technical write-up, OpenAI's account and the PHP Foundation's post · 14,358 likes and 5.73 million views on OpenAI's post at capture · October 20–21

On October 21 the PHP project shipped versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35, closing two memory-safety flaws in the engine itself 1. The flaws are CVE-2026-109618 and CVE-2026-109619, the public identifiers for the two bugs, rated 9.2 out of 10 2. Every PHP release back to 5.4 carries at least one of them 3.

Wiz Research demonstrated unauthenticated remote code execution against WordPress, Drupal and Joomla 3. It also showed an attacker reading worker memory that can hold passwords, tokens and API keys. The bugs sit below the application layer, so a plain hello-world page on an affected version is exposed too, and with PHP behind roughly 70% of websites the exposure runs through hosting providers rather than one product 34.

Calif reported an unconditional remote code execution on August 11, using GPT-5.6 Sol inside Patch the Planet, OpenAI's program for pairing models with open-source maintainers 1. OpenAI found the same root cause independently, and Wiz reported further attack vectors on August 24. Wiz credits Gemini 4 Argon with finding the vulnerability, building the exploit, writing detections and validating the patch 3.

Exposure moved fast. Cloudflare had firewall rules ready nine hours before disclosure and switched them on by default for free plans as well as paid ones 4. Its chief executive counted thousands of exploit attempts in the ten hours before the announcement, tens of thousands within 45 minutes, and millions by the next day 5.

The coordinated release did not hold. The CVE identifiers circulated on October 20 with a wrong 10.0 score, Calif published a proof-of-concept video by mistake, and Chainguard shipped patched images six hours before the agreed time 1. The PHP Foundation is blunt about the delay: competing fixes that each broke something, a security team that decides by consensus, and too few engineers who know the engine 1.

OpenAI calls these among the most impactful vulnerabilities in recent memory 6. It says frontier AI found them before attackers did, opening what the company calls a Defender's Window, the gap between a finding and its exploitation. Ethan Mollick, a Wharton professor who studies how people use AI, read the same week as acceleration of attack and defense at inhuman speed 7.

Mollick's caveat: the papers modeling this dynamic suggest AI helps attackers first, because attacks are cheap 8.

Patch PHP. A web application firewall only buys time 4, and OpenAI published a Codex skill that flags vulnerable PHP versions 9.

Claude designed 17 protein binders against scorpion toxins, and released all 211 attempts Anthropic's Science blog and technical report · 7,326 likes and 1.78 million views on the announcement at capture · October 20

Anthropic set an internal research model running inside Claude Science on a design campaign against five scorpion neurotoxins known to harm humans 10. A minibinder is a small protein built from scratch to attach tightly to a chosen target. Working from one prompt and four follow-ups, the model generated 318,484 designs with 12 design tools, co-folded 553,052 design-toxin pairs, and spent on the order of 10,000 GPU hours over 64 hours 10.

That left 180 designs, and Anthropic added 31 more with the model's help. Two contract labs, Adaptyv Bio and Twist Bioscience, tested all 211 across two expression systems, and 17 met the criteria for binding 10. Seven bind the Mexican Cn2 toxin, four AaH II, four AaH I, two LqhIII. Eleven of the 60 single-toxin designs bound, an 18% hit rate Anthropic calls typical 10.

The multitoxin goal failed. None of the designs aimed at several toxins at once bound more than one, and Anthropic lists possible causes without choosing between them: current capability, its own prompting, the limits of the design tools, or the biology 10.

The work stopped at binding assays in two labs. Turning a binder into a treatment is a separate and longer problem, and Anthropic claims a lead rather than an antivenom 10. Two outside experts, Susana Vázquez Torres and Timothy P. Jenkins, call it a promising first step and point out that binding was never the hardest part of the pipeline. All 211 designs and the lab data are public 11.

Scorpions sting about 1.2 million people a year and kill more than 3,250, mostly children, while conventional antivenom takes months of milking scorpions and immunizing horses 10. A binder designed in three days and tested in weeks changes the cost of the first step only.

Claude proved security properties in 123 open-source libraries, at the assembly level Anthropic's Frontier Red Team write-up · 6,659 likes and 986,000 views on the post at capture · October 22

Anthropic's Frontier Red Team, working with the verification company Theorem, used Claude to prove properties of compiled RISC-V code in 123 real open-source libraries 12. The authors call it the first large-scale attempt to apply formal verification to production software at the assembly level with AI.

The property proved is narrow and useful. Each program exits normally, with no memory-safety error and no failed assertion, for every input under 64 MB and every modeled Linux response and memory layout 12. The spec comes from each project's existing fuzz test, a program that throws random inputs at code looking for crashes, turned into a claim about all inputs rather than a sample.

The gain shows up in cost. The seL4 kernel took two person-years to implement and 20 to verify, and formal methods have been affordable only for mission-critical code such as Airbus avionics 12. Claude also found bugs that fuzzing had missed despite trillions of runs, and a subset of the proofs is published for anyone to check 13.

The authors state the limits themselves. The assumptions about threading, memory and the operating system bound what was shown, and the field still needs specs that describe what software is supposed to do 12. What changed is who can afford a proof.

Thousands of GPT-Rosalind agents read a mouse-brain atlas in 32 minutes OpenAI's research write-up and the paper on bioRxiv · 6,976 likes and 1.43 million views on the post at capture · October 21

OpenAI, PerturbAI and Scripps Research divided a CRISPR atlas of the mouse brain among thousands of GPT-Rosalind agents 14. Each agent took one gene disruption, read the cellular response and searched the literature. The run took 32 minutes; the team estimates the same analysis one perturbation at a time would take months to a year.

The agents surfaced a specific lead. Disrupting DCTN1, the gene behind Perry syndrome, raised the expression of cholesterol-making genes, and the agents proposed that cholesterol may not be reaching where neurons need it 14. They also found that TSC2 disruption changes dozens of communication genes in the cerebellum while TSC1 disruption barely moves them, tracking the severity gap in tuberous sclerosis.

Given the same largest workload, a single agent cited 11 references; the distributed agents cited more than 700, and blinded human experts preferred their interpretations 14. The whole atlas of genes, effects and publications is public, along with the paper 15.

Each of these is a hypothesis queued for a lab bench. OpenAI's own framing is that the output prioritizes experiments, and nothing here has been tested in tissue 14. In the same week OpenAI acquired Orchestra Bio, a platform for coordinating drug-development programs, to fold program management into Rosalind, against an average 10 to 15 years from target discovery to approval 16.

AI reviews beat expert human peer reviews in 87% of judged pairs Gary King and Steven Worthington's paper · 752 likes and 31,059 views on Mollick's post at capture · October 22–23

Gary King and Steven Worthington, of Harvard's Institute for Quantitative Social Science, put 134 pairs of reviews of the same manuscripts in front of judges 17. One review in each pair came from an AI, the other from a paid human expert whom editors and authors had rated unusually good. The corpus is 124 submissions from between 2019 and 2026, spanning statistics, formal models and qualitative work.

Human journal editors preferred the AI review in 87% of pairs, and an AI judge preferred it in 100% of the 132 pairs it decided 17. Human editors agreed with other human editors 80% of the time, so they agreed with the AI slightly more often than with each other.

The paper tested that result several ways. On a length-matched subset of 44 pairs, AI won all 44. Replicated with two model families writing and judging, AI was preferred in 97.1% of cases under an Opus judge and 94.2% under a GPT judge across 416 comparisons 17. AI reviews cited no fabricated sources, and 5.5% of their claims about the manuscript were judged incorrect against 15.9% for human reviews.

What the paper measures is the quality of review text, and its authors set out the downsides and the policy choices alongside the result 17. Ethan Mollick drew the conclusion he drew for academics generally: they need to work out how to adapt the practice of science to AI, fast 18.

OpenAI's models and Codex are coming to Google Cloud's Model Garden OpenAI's announcement · 5,104 likes and 463,000 views on the post at capture · October 21

OpenAI's frontier models and Codex agentic workflows will be available in Google Cloud's Model Garden 19. Customers buy them through the Google Cloud Marketplace, so enterprises can draw down existing Google Cloud commitments to pay for them. The pitch is aimed at regulated industries: OpenAI's models inside Google Cloud's access controls, data governance and infrastructure.

OpenAI models are also slated for Google Antigravity and the Gemini Agent, with general availability planned in the coming weeks 19. Greg Brockman calls it giving customers more choice; Thomas Kurian, Google Cloud's chief executive, calls it giving customers a broad model selection.

Google sells a rival's frontier models to keep workloads on its cloud; OpenAI reaches buyers that its own channel does not. Microsoft remains OpenAI's primary cloud partner, as the announcement itself says 19.

Two labs under outside eyes: a first public benefit report, and three weeks inside OpenAI's monitoring Anthropic's report and trustee announcement, OpenAI's post and METR's write-up · 2,504 likes and 321,000 views on Anthropic's post; 2,355 likes and 309,000 views on OpenAI's post, at capture · October 21–22

Anthropic published its first annual Public Benefit Report, measuring itself against five charter pillars and closing with its trustees' own assessment of where it is falling short 20. The same week it added Christopher Stubbs to its Long-Term Benefit Trust: an experimental physicist, Harvard's dean of science from 2018 to 2024, and a member of the team that found the universe's expansion is accelerating 21.

That trust carries real authority. It can appoint a majority of Anthropic's board, and its trustees are independent of management and investors 21. Stubbs is its first physicist and joins Neil Buddy Shah, Richard Fontaine and Ben Bernanke.

OpenAI's version of outside scrutiny was operational rather than structural. METR and Redwood Research each sent an employee to work on premises for three weeks this summer, evaluating the systems OpenAI uses to monitor its own agents for misalignment 22.

METR says the pair discovered a number of novel vulnerabilities in those monitoring systems, and OpenAI says the findings helped strengthen safeguards as it expanded the system 23. What the public gets is a brief write-up rather than the full report.